ASKII CLI ( •_•)>⌐■-■ (⌐■_■)

AI code assistant for your terminal. Powered by Ollama, LM Studio, OpenAI, Anthropic, or opencode Go.

Install

npm install -g askii-cli

Or run without installing:

npx askii-cli <command>

Interactive Mode

Run askii with no arguments to start an interactive REPL session:

askii
askii --platform anthropic   # start with a specific platform

ASKII ( •_•)>⌐■-■ (⌐■_■)  — interactive mode

  Platform : ollama (gemma4:e4b)
  Workspace: /your/project
  Wiki     : off
  Code wiki: off

Type a message to chat, /help for commands, /exit to quit.

> what does a closure do in JavaScript?

ASKII: A closure is a function that retains access to variables
from its enclosing scope even after that scope has finished...

> can you give me an example?

ASKII: Sure! Here's a classic counter example...

> /platform anthropic
Platform → anthropic (claude-sonnet-4-6)

> /do add a .gitignore for a Node.js project
[Round 1/5]
...

> /exit
Bye! ( •_•)>⌐■-■ (⌐■_■)

Bare text input maintains a persistent chat history across turns — follow-up questions remember the full conversation. Use /clear to start fresh.

REPL slash-commands

| Command | Description | | --- | --- | | /help | Show all available commands | | /ask <question> | Explicit ask (same as bare text) | | /do <task> [flags] | Run the Do agent (--max-rounds N, --yes) | | /edit --file <path> <instr> | Edit a file in place | | /explain <text> | Explain a line of code | | /wiki-reload | Rebuild the docs wiki index | | /code-wiki-reload | Rebuild the code wiki index | | /platform <name> | Switch platform for the session (also updates default model) | | /model <name> | Switch model for the session | | /config | Show current session config (keys redacted) | | /clear | Clear chat history and start a fresh conversation | | /exit, /quit | Exit interactive mode |

Tab-complete any / command by pressing Tab. Up/down arrows cycle through input history.

Config overrides — bare -- flags at the prompt update session config without restarting:

> --platform openai --model gpt-4-turbo
> --max-rounds 10
> --mode helpful

Ctrl+C during /do or /control cancels only that agent and returns to the > prompt. Ctrl+C at the idle prompt exits.

Commands

ask — Ask a question about code

Pipe code via stdin or use --code:

bash

cat myfile.ts | askii ask "what does this do?"
askii ask --code "const x = 1 + 1" "is this correct?"

PowerShell

Get-Content myfile.ts | askii ask "what does this do?"
askii ask --code "const x = 1 + 1" "is this correct?"

edit — Edit code

Returns the modified code to stdout (pipe-friendly):

bash

cat myfile.ts | askii edit "add error handling" > myfile-edited.ts
cat myfile.ts | askii edit "convert to async/await"

PowerShell

Get-Content myfile.ts | askii edit "add error handling" | Set-Content myfile-edited.ts
Get-Content myfile.ts | askii edit "convert to async/await"

explain — Explain a line of code

bash

askii explain "arr.reduce((a, b) => a + b, 0)"
cat myfile.ts | askii explain

PowerShell

askii explain "arr.reduce((a, b) => a + b, 0)"
Get-Content myfile.ts | askii explain

do — Agentic task runner

Prints the working directory's top-level file listing, then runs an agent loop that creates, modifies, renames, deletes, views, and lists files until the task is done or --max-rounds is reached.

bash

askii do "create a Jest test file for src/utils.ts"
askii do "rename all .js files to .ts in the src folder"
askii do "add a .gitignore for a Node.js project"
askii do --yes "scaffold a README for this project"   # auto-confirm all
askii do --dir ./my-project "refactor index.ts"

PowerShell

askii do "create a Jest test file for src/utils.ts"
askii do "rename all .js files to .ts in the src folder"
askii do "add a .gitignore for a Node.js project"
askii do --yes "scaffold a README for this project"   # auto-confirm all
askii do --dir .\my-project "refactor index.ts"

The agent can use the following actions each round:

| Action | Description | Requires confirmation | | ------------- | ---------------------------------------------------------------- | --------------------- | | list | List files in a folder ([file] / [folder] labels) | No | | view | Read a file's contents | No | | search | Grep workspace files for a pattern | No | | wiki_search | BM25 search over indexed .md docs (requires --use-wiki) | No | | code_search | BM25 search over indexed code files (requires --use-code-wiki) | No | | create | Create a new file | Yes | | modify | Replace text in an existing file | Yes | | rename | Rename or move a file | Yes | | delete | Delete a file | Yes | | run | Run a shell command | Yes (always) |

The loop continues after every round — not only after reads — until the AI returns [] or the round limit is hit.

control — Screen control agent

Takes a screenshot, sends it to the AI, and executes the returned mouse/keyboard action. Repeats until the AI returns DONE or --max-rounds is reached. Requires a vision-capable model (e.g. llava, moondream2).

Linux: requires xdotool for mouse/keyboard control (sudo apt install xdotool or equivalent).

bash

askii control --ollama-model llava "open Notepad and type hello world"
askii control --yes --ollama-model llava "click the search bar and search for cats"
askii control --max-rounds 10 --ollama-model llava "fill in the login form"
askii control -p lmstudio --lmstudio-model llava-1.5 "open the browser"

PowerShell

askii control --ollama-model llava "open Notepad and type hello world"
askii control --yes --ollama-model llava "click the search bar and search for cats"
askii control --max-rounds 10 --ollama-model llava "fill in the login form"
askii control -p lmstudio --lmstudio-model llava-1.5 "open the browser"

Each round the AI can return one of:

• mouse_move — move the cursor to (x, y)
• mouse_left_click — left-click at (x, y)
• mouse_right_click — right-click at (x, y)
• keyboard_input — type a string
• DONE — task complete, stop the loop

Without --yes, each proposed action is shown with its reasoning and requires y confirmation before executing.

wiki-reload — Index wiki documentation

Walks all .md files under --wiki-path, splits them into sections by heading, builds a MiniSearch BM25 index, and saves it as .askii-wiki-index.json inside the wiki folder. Run this once after pointing --wiki-path at your docs, and again whenever the docs change.

bash

askii wiki-reload --wiki-path ./docs
askii wiki-reload --wiki-path /home/user/my-project/docs

PowerShell

askii wiki-reload --wiki-path .\docs
askii wiki-reload --wiki-path C:\my-project\docs

After indexing, pass --wiki-path and --use-wiki to any ask, edit, or do command to inject the top matching documentation chunks as context:

bash

askii ask --wiki-path ./docs --use-wiki "how do I configure the database?"
cat src/db.ts | askii edit --wiki-path ./docs --use-wiki "add connection pooling"
askii do --wiki-path ./docs --use-wiki "implement the auth flow described in the docs"

PowerShell

askii ask --wiki-path .\docs --use-wiki "how do I configure the database?"
Get-Content src\db.ts | askii edit --wiki-path .\docs --use-wiki "add connection pooling"
askii do --wiki-path .\docs --use-wiki "implement the auth flow described in the docs"

code-wiki-reload — Index workspace code files

Walks all supported code files in a directory (TypeScript, Python, Go, Rust, C/C++, and more), splits them into 60-line overlapping chunks, builds a MiniSearch BM25 index, and saves it as .askii-code-wiki-index.json in the target directory. Run this once against your codebase and again whenever the code changes significantly.

Skips: node_modules, dist, out, build, target, bin, obj, coverage, __pycache__, venv, and files over 200 KB.

bash

askii code-wiki-reload                          # indexes current directory
askii code-wiki-reload --code-wiki-path ./src  # indexes a specific directory

PowerShell

askii code-wiki-reload
askii code-wiki-reload --code-wiki-path .\src

After indexing, pass --code-wiki-path and --use-code-wiki to any ask, edit, or do command to inject the top matching code chunks as context:

bash

askii ask --use-code-wiki "where is the wiki index built?"
askii ask --use-code-wiki "how does the debounce work in the inline completer?"
cat src/commands.ts | askii edit --use-code-wiki "add code wiki context to the ask command"
askii do --use-code-wiki "refactor the provider functions to share a common helper"

PowerShell

askii ask --use-code-wiki "where is the wiki index built?"
Get-Content src\commands.ts | askii edit --use-code-wiki "add code wiki context to the ask command"
askii do --use-code-wiki "refactor the provider functions to share a common helper"

browse — Browser agent

Launches a Puppeteer browser, takes a screenshot of the current page and its URL, sends both to the AI, and executes the returned action. Repeats until the AI returns DONE or --max-rounds is reached. Requires a vision-capable model (e.g. llava, moondream2).

By default the browser window is visible. Pass --headless to run in the background.

Requires Chrome or Chromium to be installed. Use --chrome-path (or ASKII_CHROME_PATH) to specify the executable path if it is not detected automatically.

bash

askii browse --ollama-model llava "go to https://example.com and click Learn more"
askii browse --yes --ollama-model llava "search Google for Node.js and open the first result"
askii browse --headless --yes --ollama-model llava "check the title of https://github.com"
askii browse --max-rounds 10 --ollama-model llava "fill in the login form on example.com"
askii browse -p lmstudio --lmstudio-model llava-1.5 "go to news.ycombinator.com"
askii browse --chrome-path "/usr/bin/chromium" --ollama-model llava "go to example.com"

PowerShell

askii browse --ollama-model llava "go to https://example.com and click Learn more"
askii browse --yes --ollama-model llava "search Google for Node.js and open the first result"
askii browse --headless --yes --ollama-model llava "check the title of https://github.com"
askii browse --max-rounds 10 --ollama-model llava "fill in the login form on example.com"
askii browse -p lmstudio --lmstudio-model llava-1.5 "go to news.ycombinator.com"
askii browse --chrome-path "C:\Program Files\Google\Chrome\Application\chrome.exe" --ollama-model llava "go to example.com"

Each round the AI can return one of:

• goto — navigate to a URL
• click — click an element by CSS selector
• type — type text into an element by CSS selector (clears existing value first)
• wait_for — wait until a CSS selector appears in the DOM
• back — navigate back in browser history
• forward — navigate forward in browser history
• DONE — task complete, stop the loop

Without --yes, each proposed action is shown with its reasoning and requires y confirmation before executing.

Options

| Flag | Short | Description | Default | | ------------------- | ----- | ---------------------------------------------------------------------------------- | ------------------------ | | --platform | -p | LLM platform: ollama, lmstudio, openai, anthropic, opencodego | ollama | | --ollama-url | | Ollama server URL | http://localhost:11434 | | --lmstudio-url | | LM Studio server URL | ws://localhost:1234 | | --ollama-model | | Ollama model | gemma4:e4b | | --lmstudio-model | | LM Studio model | qwen/qwen3-coder-30b | | --openai-key | | OpenAI API key (env: ASKII_OPENAI_KEY) | | | --openai-model | | OpenAI model | gpt-5-mini | | --openai-url | | OpenAI-compatible base URL (env: ASKII_OPENAI_URL) | | | --anthropic-key | | Anthropic API key (env: ASKII_ANTHROPIC_KEY) | | | --anthropic-model | | Anthropic model (env: ASKII_ANTHROPIC_MODEL) | claude-sonnet-4-6 | | --opencodego-key | | opencode Go API key (env: ASKII_OPENCODEGO_KEY) | | | --opencodego-model| | opencode Go model (env: ASKII_OPENCODEGO_MODEL) | glm-5.2 | | --opencodego-url | | opencode Go base URL (env: ASKII_OPENCODEGO_URL) | https://opencode.ai/zen/go/v1 | | --mode | | Response style: helpful, funny | funny | | --max-rounds | | Max agent rounds for do / control / browse | 5 | | --dir | | Working directory for do | cwd | | --code | -c | Code input (alternative to stdin) | | | --yes | -y | Auto-confirm all actions | | | --headless | | Run Puppeteer headlessly for browse | false (visible) | | --chrome-path | | Path to Chrome/Chromium executable for browse | | | --wiki-path | | Path to folder with .md docs for wiki RAG (env: ASKII_WIKI_PATH) | | | --use-wiki | | Inject wiki context into ask / edit / do (env: ASKII_USE_WIKI=1) | | | --code-wiki-path | | Path to codebase root to index / search (env: ASKII_CODE_WIKI_PATH) | cwd | | --use-code-wiki | | Inject code wiki context into ask / edit / do (env: ASKII_USE_CODE_WIKI=1) | |

Environment Variables

bash

export ASKII_PLATFORM=ollama

# Ollama
export ASKII_OLLAMA_URL=http://localhost:11434
export ASKII_OLLAMA_MODEL=gemma4:e4b

# LM Studio
export ASKII_LMSTUDIO_URL=ws://localhost:1234
export ASKII_LMSTUDIO_MODEL=qwen/qwen3-coder-30b

# OpenAI
export ASKII_OPENAI_KEY=sk-...
export ASKII_OPENAI_MODEL=gpt-5-mini
export ASKII_OPENAI_URL=   # leave empty for api.openai.com

# Anthropic
export ASKII_ANTHROPIC_KEY=sk-ant-...
export ASKII_ANTHROPIC_MODEL=claude-sonnet-4-6

# opencode Go
export ASKII_OPENCODEGO_KEY=...
export ASKII_OPENCODEGO_MODEL=glm-5.2
export ASKII_OPENCODEGO_URL=https://opencode.ai/zen/go/v1

# Shared
export ASKII_MODE=funny
export ASKII_MAX_ROUNDS=5
export ASKII_CHROME_PATH=/usr/bin/chromium

# Docs wiki RAG
export ASKII_WIKI_PATH=./docs
export ASKII_USE_WIKI=1

# Code wiki RAG
export ASKII_CODE_WIKI_PATH=./src   # defaults to cwd
export ASKII_USE_CODE_WIKI=1

PowerShell

$env:ASKII_PLATFORM = "ollama"

# Ollama
$env:ASKII_OLLAMA_URL = "http://localhost:11434"
$env:ASKII_OLLAMA_MODEL = "gemma4:e4b"

# LM Studio
$env:ASKII_LMSTUDIO_URL = "ws://localhost:1234"
$env:ASKII_LMSTUDIO_MODEL = "qwen/qwen3-coder-30b"

# OpenAI
$env:ASKII_OPENAI_KEY = "sk-..."
$env:ASKII_OPENAI_MODEL = "gpt-5-mini"
$env:ASKII_OPENAI_URL = ""   # leave empty for api.openai.com

# Anthropic
$env:ASKII_ANTHROPIC_KEY = "sk-ant-..."
$env:ASKII_ANTHROPIC_MODEL = "claude-sonnet-4-6"

# opencode Go
$env:ASKII_OPENCODEGO_KEY = "..."
$env:ASKII_OPENCODEGO_MODEL = "glm-5.2"
$env:ASKII_OPENCODEGO_URL = "https://opencode.ai/zen/go/v1"

# Shared
$env:ASKII_MODE = "funny"
$env:ASKII_MAX_ROUNDS = "5"
$env:ASKII_CHROME_PATH = "C:\Program Files\Google\Chrome\Application\chrome.exe"

# Docs wiki RAG
$env:ASKII_WIKI_PATH = ".\docs"
$env:ASKII_USE_WIKI = "1"

# Code wiki RAG
$env:ASKII_CODE_WIKI_PATH = ".\src"   # defaults to cwd
$env:ASKII_USE_CODE_WIKI = "1"

Platforms

Ollama (default)

bash

ollama pull gemma4:e4b
askii ask "what is a closure?"

PowerShell

ollama pull gemma4:e4b
askii ask "what is a closure?"

LM Studio

bash

# Start LM Studio with local server enabled
askii -p lmstudio ask "explain this function"
askii -p lmstudio --lmstudio-model "my-model" ask "explain this function"

PowerShell

# Start LM Studio with local server enabled
askii -p lmstudio ask "explain this function"
askii -p lmstudio --lmstudio-model "my-model" ask "explain this function"

OpenAI

bash

askii -p openai --openai-key sk-... ask "what does this do?"
askii -p openai --openai-key sk-... --openai-model gpt-4-turbo do "add error handling"
# Azure OpenAI or any compatible API:
askii -p openai --openai-key sk-... --openai-url https://my-resource.openai.azure.com ask "explain"

PowerShell

askii -p openai --openai-key sk-... ask "what does this do?"
askii -p openai --openai-key sk-... --openai-model gpt-4-turbo do "add error handling"
# Azure OpenAI or any compatible API:
askii -p openai --openai-key sk-... --openai-url https://my-resource.openai.azure.com ask "explain"

Anthropic

bash

askii -p anthropic --anthropic-key sk-ant-... ask "what does this do?"
askii -p anthropic --anthropic-key sk-ant-... --anthropic-model claude-sonnet-4-6 do "add error handling"
askii -p anthropic --anthropic-key sk-ant-... --anthropic-model claude-haiku-4-5 explain "arr.reduce((a, b) => a + b, 0)"

PowerShell

askii -p anthropic --anthropic-key sk-ant-... ask "what does this do?"
askii -p anthropic --anthropic-key sk-ant-... --anthropic-model claude-sonnet-4-6 do "add error handling"
askii -p anthropic --anthropic-key sk-ant-... --anthropic-model claude-haiku-4-5 explain "arr.reduce((a, b) => a + b, 0)"

opencode Go

A hosted, multi-model coding subscription (opencode.ai/go). Most models use an OpenAI-compatible endpoint; Qwen and MiniMax models use an Anthropic-compatible one — ASKII routes automatically based on the model id. See the full model list at opencode.ai/zen/go/v1/models.

bash

askii -p opencodego --opencodego-key ... ask "what does this do?"
askii -p opencodego --opencodego-key ... --opencodego-model kimi-k2.7-code do "add error handling"
askii -p opencodego --opencodego-key ... --opencodego-model qwen3.7-max explain "arr.reduce((a, b) => a + b, 0)"

PowerShell

askii -p opencodego --opencodego-key ... ask "what does this do?"
askii -p opencodego --opencodego-key ... --opencodego-model kimi-k2.7-code do "add error handling"
askii -p opencodego --opencodego-key ... --opencodego-model qwen3.7-max explain "arr.reduce((a, b) => a + b, 0)"